Generative Sequence Modeling for Maritime Behaviors
Generative models for maritime trajectory forecasting and imputationGenerative Sequence Modeling for Maritime Behaviors builds probabilistic and deep generative models to address missing data, irregular sampling, and future trajectory prediction in maritime sensor streams.
Key Contributions
Missing Data Reconstruction: Real-world maritime streams often contain dropouts and irregular sampling, which degrade early-intent performance. CVAE-based latent models and LSTM-GAN generators learn the joint distribution of past and future motion, serving as imputers for incomplete trajectories.
Future Trajectory Prediction: Models simulate multiple plausible futures, supporting scenario forecasting, counterfactual analysis, and data augmentation for adversarial events.
Multi-Task Design: Designed architectures that jointly predict future vessel motion and latent intent, enabling richer representations for downstream decision systems.
Integration with NavySim: Models integrate with the naval simulation platform for scenario generation and evaluation under uncertainty.
Research Directions
- CVAE, TimeGAN, ACGAN, and LSTM-GAN–inspired variants
- Encoder–decoder models for long-horizon forecasting
- Simulation of adversarial naval behaviors
